I would be really curious to hear from devs at Databricks what the experience of development is like internally. I work at a small startup with essentially unlimited AI spend budget - the entire point is that I should be turning to it at every opportunity since our human labor is so expensive relative to tokens. So generally it's like:
- Spend most time prioritizing/discussing what to do.
- Once that's agreed, use Fable 5 High + 5.6 Sol XHigh come up with a design + plan. Agree on the high level plan. (Usually this just comes down to choosing where the change belongs on the spectrum between minimal patch <-> full redesign)
- Use Opus 5 or Sol Med to execute
- Auto-fix bugs and CI until green + thermonuclear review skill x3.
- Manual interrogation of change/nits
- Come up with QA plan and have Codex Computer Use execute on it
- Manually spot check the final result (usually a sizable diff, thousands of lines, complete feature E2E, etc)
I probably spend like $80 a day at least but I produce the output of 3 or 4 2022 engineers and probably at better quality. So it's easily worth it. Would I save money by switching to GLM 5.2 and such...perhaps? IDK. At our scale it's not worth the time spent building the eval harness to actually understand the performance tradeoff.
I'm also at a startup. My workflow is similar but I have Fable 5 xhigh drive the whole thing: it gets Codex CLI installed in its environment with an API key, and it's instructed to delegate ~everything to Codex and review its work, especially for code quality/conciseness. Fable delegates to Sol or Luna (fast mode) xhigh/max depending on the task - I think Luna xhigh on fast mode is basically a Pareto improvement over Sol medium.
In our team's experience, the product of agents is generally The Homer (1). It does work, but it's vastly overengineered.
When I personally want tight code, I have to spend a considerable amount of time adjusting it manually:
- It needs to be trimmed down. In my experience, at least one agent I use struggles to produce minimalist designs, and it's very frustrating
- I need to consider whether there are solutions based on higher-level assumptions, that AIs typically miss
- I need to check whether there are off-the-shelf solutions - AIs like to reinvent the wheel
IMO, software production has become a mass-produced commodity in every sense - it's much more expensive to produce software manually, but the quality is not the same.
It took me a few seconds of deliberating if The Homer was a reference to baseball or "The Odyssey" and then realized there was a footnote
> software production has become a mass-produced commodity in every sense - it's much more expensive to produce software manually, but the quality is not the same.
I believe “The Homer” is a reference to Season 2, Episode 15 of The Simpsons, “Oh Brother, Where Art Thou?”
Homer reunites with his long lost brother, who runs car company Powell Motors. Homer is ultimately tasked by his brother with helping to design a car for the “average man” that ultimately bankrupts the company for being wildly overengineered and costing too much ($82,000 in 1991-money).
Since software is still a winner-takes-all market, the mass-production property of software doesn't really matter.
In such markets, what you produce is either worth nothing or worth millions of dollars. For as long as it's the case that well-constructed code (with or without LLM help) is more likely to be in the latter category, the economics of software don't really change.
Even before LLMs, you could've commissioned a half-assed clone of any app you wanted from a 3rd world consultancy for a few thousand dollars. LLMs are basically Bangalore-as-API.
I think part of the reason software was winner take most was the difficulty of making software.
I remember hearing a story that in the past movies were so technically difficult to make that any movie that got made had a good chance to be a profitable hit. But as movies got cheaper to make, more movies got made. Nowadays movie studio execs have to really calculate out the audience and expected revenue for any new movie and balance that against the budget and the cost of the studio's failed movies.
> I think part of the reason software was winner take most was the difficulty of making software.
That might be part of it, but I think it also has to do with the reality of replicating and scaling. Hardware or physical goods simply don’t scale like digital goods. There can be hundreds of knock-off physical products that have lower quality and lower cost but serve 90% of the same purpose, because physical capacity for raw materials, construction, labor, shipping, etc. have scaling limits in each market and economy. Digital goods are just so much easier to replicate and scale, so it often doesn’t make sense to buy software at lower quality and lower price if it doesn’t do most of the job. There are still limits of course, and different from physical goods, but I think this is a key reason why software is seen as winner-take-all.
That honestly sounds a lot healthier than “just ship what the CTO/Product team wants” with as much hand waving as is necessary to very roughly estimate ROI and then pray it hits with the market. In anything that’s not a startup operating in a new industry, the “old way” is a hard way to run a business
> Since software is still a winner-takes-all market, the mass-production property of software doesn't really matter.
Fairly sure software dev was always an iceberg. Most software and most software devs aren't working on horizontal software, but on vertical software, in cost centers. Sales for that kind of software don't scale as much.
No surprise, LLM companies optimize for waste. More tokens, and more prompts means more revenue. Reminds of Google’s Prabhakar Raghavan story: deliberately making search worse [1]
Or, more likely, it's that concise code requires a much deeper, wholistic, understanding that these models just are capable of yet.
Same with a junior dev. They don't write long form spaghetti because they're trying to write more LOC. They do it because not doing it is hard, literally above their pay grade.
I use LLM every day, but they're still completely awful at architecture. I don't think this clear lack of ability is some conspiracy.
> software production has become a mass-produced commodity
For who?
The public? The public has never liked buying software at any price.
Businesses? Businesses need higher quality software when it's relevant to their core competencies, so they hire people instead. Buying competing SaaS or depending too much on AI is throwing the baby out with the bathwater.
This was more true a few months ago but Fable has improved the situation considerably.
Also just remember - minimalist code looks and feels great but customers do not read your code. I have caught myself many times providing "corrections" to abstractions that were already ~fine, just not perfect. The average SWE costs $200/hr. Careful you don't burn $50 worrying about code that will likely be rewritten or can be better abstracted when that's actually needed.
This is a pointless quibble but the hourly rate claim is not true--it's like ~$60 in the USA [0]. Maybe you meant at a specific Org but this is important context when comparing "pricing" between human and AI.
How is this not true? Taking a Senior SWE @ ~$200K, even just the base salary cost / 2080 working hours is $100/hr. Fully loaded employer cost + accounting for non-coding time gets you to upper 100s easily.
Even for a junior making $100K, I have a hard time believe their time is worth less than $75/hr or so.
Edit: Fine, "Senior" is not "Average". But naive salary is not the true numerator.
Western Europe is mostly consultancy, and the rate paid by client is usually higher, and doesn't matter if it's eastern Europe, Portugal or even India.
It is. And the quality is on par with any us eng. People here forget that the big comp packages are a minority even in the US. The cost tho is much higher than just salary.
I hire contractors for a large enterprise in the US. The going rate is typically $85-$100/hr for a senior dev, depending on specialization. Lead-level maybe $120 for the right skill set.
Of course, the SWEs making that much (over 200k) are not representative of the broader field. That's the point.
Pay hits a ceiling, and that ceiling is moving lower regardless of experience. That has nothing to do with AI, but what the market will bear. Hiring counts of humans must increase no matter what. Moving some of the spend to AI reduces the risk of hiring less qualified employees they might have rejected a decade ago.
Wages at the top end are stagnating to subsidize this. That's undeniable.
An MBA's rule of thumb is that a full time employee's hourly cost to a business is at least 1.5x to 2x times their salary depending on employer taxes, benefits, offices, travel, training, hardware, perks, etc.
Minimalist code is necessary to keep AI agents working well for longer than a month on a system IME. At a certain point, their own machinations overwhelm them and they both slow down, and make worse and worse decisions.
"Will I benefit from this code being minimalist before [date]", where [date] is whenever you think the agent will be good enough to come back and make the corrections you would make today.
I'd caution that some corrections become harder to make over time, rather than easier. A bad architecture now can become much harder to fix once other things have grown up around it.
Even as a self-contained unit, you can't step in the same river twice, and on [date] some important details may have seriously faded, both in terms of text that can be mined and also in terms of human "why did we do that" and "what was the reason we did it this way and not that way" etc.
That "cost" includes all the overhead provided by the company: benefits, rent for offices, utilities, equipment, etc. The average SWE is not taking home anything close to that, outside of Silicon Valley and a few other limited areas.
Whatever you are making this year as SWE you'll be making less next year if the current trend in improvement of AI coding aids is going to be sustained. Think about it : programmers used to derive a lot of their value from the fact that it was a hard skill to acquire. My kids can now 'vibe code' stuff faster (and better looking) than what I could come up with as the beginnings of a design plan. And then I still need to implement it.
There's a massive difference between your kids vibe coding something and an engineer using AI to implement something. If you're unable to discern the difference, that's something to reflect on :)
CEOs will always want someone who knows to implement so we're safe from kids vibe coding their way in but in a short while it becomes AI who knows who is managing less expensive AI.
IME this works until it does not. This approach works well at the beginning of a greenfield project, but at the same time because it is so easy to add features, you will likely ship something that is way too over engineered. And that complexity will not amortize over next increments and will more likely lead to the entire project being a black box only fully understood by AI. However a more careful use of AI for targeted surgical changes is far more ”productive” in the long term IMO.
I disagree, the approach works well in a legacy project, since there are structures and standards that already exist, which them model can draw from (if you aren't more explicit about it in AGENTS.md)
This! I don't think folks understand how easy it is to go from greenfield to brownfield with these tools, esp if your organization is only valuing velocity. Meaning your doing full agentic development on large features, barely reviewing any code, and shipping without much refinement. It's insane, but this appears to be the status quo in SF startups.
Something isn't clear about the size of your codebase here and the level of reliability your customers expect, as a reader of your comments. Clarity there will help.
My observation has been:
- Initial greenfield work by an LLM is fast and very effective with minimal or no human oversight.
- Subsequent work ends up being over engineered and very verbose. Assumptions are made that aren't suited to the problem at hand (for example I find Fable is extremely regex happy where structured data would work much better from a readability perspective.)
- Once code bloats beyond a certain point due to unguided LLM usage, complexity is high enough that only LLMs can operate on the codebase with any economical amount of time.
- Rinse repeat and your code ends up unclear about any state that's not explicitly being tested and verified in QA loops
For some of our products this has been fine, for others it's been problematic. An understanding of your size and reliability requirements will help make the conversation more productive.
Why aren't you guiding your LLM usage? Is that what I said - to spam it and not guide anything? Or to have a careful workflow where you agree on design and maximize your human judgement/leverage?
> any state that's not explicitly being tested and verified in QA loops
As opposed to before, when engineers perfectly reasoned about code behavior from first principals and QA was unnecessary?
> Why aren't you guiding your LLM usage? Is that what I said - to spam it and not guide anything? Or to have a careful workflow where you agree on design and maximize your human judgement/leverage?
You didn't say anything positively or negatively regarding this so I made an assumption that you were using the LLM relatively unguided (e.g. a bit of oversight, not the kind of thing that heavy code reviews used to involve pre-agents.) Feel free to add clarity on your actual usage loop.
> As opposed to before, when engineers perfectly reasoned about code behavior from first principals and QA was unnecessary?
In my experience, most engineers are quite good at reasoning about code behavior for non-QAed code paths. Obviously things fall through the cracks. But I've been in the ground floor of plenty of Big Techs in their early stages before agents and, yes, a lot of initial development had spotty test coverage and yet most of the engineers had good mental models of what was happening. It used to be a very valuable skill to wrap your head around a torrid piece of code with few or no tests but was nonetheless a core piece of your application. Conversely, agentic development can bring cognitive debt [1].
===
This isn't a fight. We aren't sparring over what's right and wrong. I'm just curious how other people use agents in their work as someone who is also now in a startup that uses LLM agents heavily and has no limitations on spend.
> You didn't say anything positively or negatively regarding this so I made an assumption that you were using the LLM relatively unguided
I feel like this statement betrays your lack of advanced experience coding with LLMs.
OP's elaboration of the steps they are going through (planning, agreeing on plan, getting one LLM to draft execution plan, approving it, then executing with a separate LLM, then reviewing/testing) made it super obvious to me that they are guiding their LLMs quite considerably as part of their work.
Anyone making blanket statements about LLMs producing garbage is just telling on themselves about not having proper SDLC practices in place.
Planning, agreeing on a plan, separating planning and implementation LLM, using separate review LLMs, these are all table stakes. This isn't "guidance" if you're getting paid to write software. If you think "unguided" means "I typed a prompt into claude code and waited yolo" I don't know what to say but, you have a very different idea of what professionals do than I do.
I find for my own work that I need to read the diff the LLM produces then offer feedback on the diff in its own loop before I am satisfied, and this is after all the unattended QA steps through Codex Computer or Claude MCPs happen. Then auto reviewers come in and then reviewers come in. Of course, at our stage, we rarely have this luxury and it's only reserved for the very core of our codebase.
This is still much less guidance than we used to do for code before agents became popular. Even at Series A companies, before agents, we used to socialize tech specs, get buy-in from multiple engineers, create test plans, etc etc.
> Anyone making blanket statements about LLMs producing garbage is just telling on themselves about not having proper SDLC practices in place.
> I feel like this statement betrays your lack of advanced experience coding with LLMs.
Are we in school debate club? I don't know what's going on lol, I'm just curious how people are using LLMs! Is it just that irresistable to take a cheap shot at each other?
Not that I know of but that's the conclusion I drew from your statement.
It's not a cheap shot unless you took it personally?
I suppose I could have said "the fact that OP's explanation of how they work did not lead you to conclude they were in fact guiding their LLM usage quite a bit tells me that perhaps you have not been working with LLMs in any advanced capacity".
For the SDLC comment I admit it was a broader statement (based on observing people generalizing that "LLMs produce bad outputs") and not specifically aimed at you, and I didn't make that clear, so my bad.
> If you think "unguided" means "I typed a prompt into claude code and waited yolo" I don't know what to say but, you have a very different idea of what professionals do than I do.
AI is an accelerate tool for any organizations, management thinks it'll solve their organization issue because it accelerates it. Most often, it accelerates toward a wall.
Design is too expensive, we do agile.
QA too expensive, we fire all of them, and claim devops is the now, which allows us to fire the Ops team too, 100% ownership from deisng to ops on devs.
One person with an agent can replace all these teams. Yeah mo profits.
What is the point of working at a startup if you’re dealing with millions of lines of legacy code ? Isn’t the whole point of startups to create & innovate with a clean slate and modern tools?
How long has your startup been around? I’ve worked at plenty of startups over the past 20 years. Including one that was still calling themselves a startup 10 years out. The org I work at now was a startup before my tech giant employer acquired them. We have a very bloated and very profitable 8 year old codebase that is barely 500k LOC.
I’ve never seen a startup with a multi million line legacy codebase.
In response to a question about a legacy codebase at a startup. That implies that they think whatever they are doing is common. And forking a multi million line codebase and heavily developing it isn’t common for startups.
I don't know if that's the whole point, but I agree with the sentiment, why would a startup be working in legacy code and where would that code come from if this is truly the start of something.
OP might just be working at a small software company or for one that broke from a bigger one and is now "startup" like?
You'd be surprised. I met a guy last week who was proud to tell me he had vibe coded an almost 2 million line code base. The app did not sound that complicated, so I'm assuming it's full of copy-pasta flavored slop.
"startup" and "legacy codebase" are diametrically opposed concepts.
And if you're saying (based on your other comments) that a 6 month window is enough to create a legacy codebase...that indicates a serious lack of experience or understanding as to what a legacy codebase is, or why they exist.
Man, so many people in this thread just arguing pointless semantics, making weirdo absolutist (and incorrect) statements.
Accept that other people may ascribe different meanings/interpretations to words than you, and that if your reading of their statement doesn't make sense to you, perhaps you are simply reading it wrong.
Trying to hold someone else to your definition of words suits what purpose exactly? Are you just trying to "win" ?
Yes lol. Of all things people are getting on me for it's the number of LoC x Years In Business of this startup. I don't fucking know, I didn't start the company and I wasn't here for several of those industrious years. Looking now it looks like we have slightly fewer LoC than that, I was counting some of the generated stuff.
But who cares? The point is any codebase over a few years old with lots of customers and a big surface area has lots of code, much of it "legacy" from the standpoint of a guy in 2026.
This is an ultra cop out. There are standards in language that are not all “left means right for me so you cannot assume when i say right it is right and not left”
This whole thread around loc is depressing. It speaks volumes of some peoples inexperience working on actual legacy code. Legacy code is not just age or size but that the technical foundation is dated in a fundamental way. A giant monolith running on a now defunk framework using a database only one guy in canada knows about.
Case and point in my day job. The org that owns XMM development does not know how to recover a physical bench that is bricked because everyone who knew how has left. So now they just use simulators…
Interestingly, AI figured out some of this pretty easily for me. But the org has the exact same AI as i do. At the same time another org is close to a year into a greenfield rewrite that has been developed via agentic swarms. Absolute trainwreck.
AI doesnt make bad engineers good. Anyone who says they are doing 4 eng work likely would be without ai too. Those that claim otherwise, are the bad engineers.
exactly. Usually legacy code forms when people lose context and confidence in parts of the codebase due to staff turnover etc and ppl avoid touching or enhancing those parts for long periods. Six months is a short time to accrue that much tech debt, its enough time where most of the people who created that "legacy" are probably still around. As you said indicates bigger problems.
I'm probably between $50-$200/day depending on the day; we also have effectively unlimited budget, though a lot of that is because Azure gives startups $150,000 in credits for 2 years, which we've wired up to a LiteLLM gateway & OpenCode. Without that I think our appetite would be more around $400/month/employee.
A lot of my high costs is because I just throw Sol at everything. If I were more selective and brought in Luna or v4 Flash every once in a while, I think I'd be more like ~$400/month. That's why I'm not aligned with the notion that "tokens are subsidized so that's why people are using so much": its not that I'll have to adjust to using less, its just that I'd need to think before I prompt a bit and be more judicious. I could easily see my raw token counts doubling or tripling in the coming months. I don't think that will change as subsidization subsides; though maybe lab revenue will; intelligence per dollar is getting cheaper every week. Its solely a function of adaptation to process, which takes time.
The productivity gains per token are the single most asymmetrical thing I've ever seen in engineering. The engineers on our team are pretty effective with tokens; easily that 2x-4x output as you're seeing, spending $20-$200/day. Some of our security folks have also started contributing more-and-more code, and they're on the other side: they'll spend hundreds a day running in circles, eventually producing these +/-30k loc pull requests that take ages to get merged and are littered with issues. They weren't writing much code before, so arguably they're more productive by some multiplier greater than 1, but I think the drag on the rest of the team, and potential issues with what they produce, has overall created a net-negative situation. Inversely, some other company functions have produced a few one-off websites for things like sales processes, and those have been a huge win. The asymmetry is wild. There's almost a valley of incoming skill where if you know nothing about code, you'll leverage it well; if you know just a little bit, it makes you super dangerous; if you know a lot, you're the biggest winner. Really difficult situation to navigate.
Do you have issues with performance at the moment? Right now I tend to find that it produces absolutely terrible design patterns and especially performance. I mean maybe I don't know exactly what area you're looking at but yeah for us we tend to find it's terrible wrt dB/caching/scaling and often any performance improvements it proposes end up actually shooting itself in the foot and being worse than before but it's not very good at testing in an organized way to even notice it made it worse despite repeated prompts to do so I mean if I prompt it to test performance in a handheld structured way (it is very bad at finding out what performance to test and why) before making changes I can usually figure it out but it usually takes insistence on the specifics to really ensure a good solution that will actually fix the problem
Performance is better than ever. It's never been more practical to set up wildly complex synthetic test environments and measure perf wins. Plus the models will find every possible algorithmic/design improvement.
It actually gives me quite an uncanny feeling, bulldozing over years of human optimization work with a newer, "perfect" design. Like bringing an AK-47 back to the middle ages.
> the models will find every possible algorithmic/design improvement
it's so hard to square such totalizing statements with my day to day experience with fable and sol, (every possible, improvement, really?? they are NOT omniscient) arguing with them/my colleagues' agents that no they have slowed down the system 200x with their terrible change, doing string operations on millions of db rows, trying to get it to understand that I don't care that it's calling it a "cache" if a cache hit is slower than what we had before.
These agents do let you learn codebases quickly, and produce code way faster. I don't look at IDEs all that often. But literally multiple times every single day I catch them doing something stupid.
I don't think its impossible that we could get better performance from the agents. I know ive tried all sorts of workflows and skills, few of which seem to have much effect on the things the models struggle with. I think a big part of it is encoding enough context for large codebases, and providing it with all the tools it needs to make it successful, things to automatically check its work, etc. But that's not automatic, in fact its generally a terrible judge of what it needs or what its bad at
yeah it's true, you do have to guide them. i find that the key is you have to know what's possible. you have to have the instinct for "this really shouldn't be so difficult". my junior SWE coworkers have the same trouble as your coworkers.
but the revolution is it doesn't take that long. in like 15 minutes you can chat with fable and get to the meat of whatever the issue is with repeated questioning. and then it does the solution for you. so it's not magic but it's still like a 100x speedup.
I encounter this regularly and it still feels weird.
That sense that you did something better in a few days than you would have in a month 5 years ago. It's like buying a table saw for wood working.
One crazy thing I think about often is how there are so many correctness and testing harnesses that would have taken weeks to build in the past so we simply never would have. We'd just do our best then wait and see what comes to the surface. This is a huge part of what makes it possible to actually make better software with LLMs in my opinion. It isn't just 'LLM codes better than I ever could' (that's often untrue still) but 'LLM enables me to make assertions about the program to degrees that would have been absurdly impractical in the past'. It's huge
Yes 100%. This morning I casually prompted Codex to drive the browser to complete extensive performance testing in-situ that would have literally been weeks of work before. Probably in reality it just wouldn't have been done, and performance guarantees would have been attempted up front via more careful design.
In this case the design was also AI generated, and there were limited wins to be found because the design was already superb.
Are you using the SOTA models at very high reasoning during planning? IME that makes a LOT of a difference. I‘d also never let them just rip into the architecture, but always push back and ask for alternatives first. Once the overall plan is nailed, not that much can go wrong. Provided it’s a reasonable change set and not a 20k LOC PR.
fable or sol w/ very high both planning and execution, yeah. I feel the "push back" part is a big part of my job now (on every step, planning, execution, and review) yeah, but that feels pretty incompatible with the sorts of "just let it do what it wants" which other people seem to be claiming
> Spend most time prioritizing/discussing what to do.
you should probably be doing this discussion work along with Fable 5. It will give good feedback if you're working on the correct things.
> Come up with QA plan and have Codex Computer Use execute on it
QA plan should be part of the above "design + plan", not after it. The implementer needs to be able to fully test before publishing a PR. This is true whether humans or agents are writing the code.
> Manually spot check the final result (usually a sizable diff, thousands of lines, complete feature E2E, etc)
unfortunately this is not really scalable with amount of code agents can produce, so you need independent (fresh context) agent reviewers to help. Ideally they only escalate to a human when really stuck.
> I probably spend like $80 a day at least
at a small startup you should be on the $200/month plan(s).
$80 is definitely low now that I look at my numbers. but not OOMs low, it's closer to like $200 on heavy days. i don't know how you're doing $3k/day, that's wild. i'm pretty aggressive about compaction and session restarts, and i reserve Fable 5/Sol XHigh for "main thread" orchestration
Only spending $80 a day on Opus 5/Fable 5/GPT 5.6 Sol feels very low. I'll roll through a couple hundred dollars worth of credits a day with those models, the vast majority of which would be on non-coding tasks, and it's still a huge cost savings over me or my team having to do these things manually, if we'd even be able to do them at all.
But that's also why it's now easy to justify the cost of an Nvidia or Intel inference server with Kimi K3 locked and loaded :)
The output yes, but do you produce the impact and value of 3 engineers? I have seen this workflow being toyed with too, and I find it to produce massively overengineered stuff that actual people don't really wanna use
My experience is that your description works for a certain time, since you're knowledgeable of the codebase and can guide it. But after too many iterations with not hand-holding the llm, it quickly gets unwieldy.
Are you at least conversational in the subject matter? You're gonna have a good time just by paying attention and adjusting your workflow. If you're getting a lot of back and forth with it, its asking a lot of planning type questions, stop, step back, rethink the whole feature, and start again from the beginning with everything more fleshed out.
If you are in a brand new field, there's no way to bridge that divide. The issue is you don't know what is good or bad, or whether what you have learned is good or bad. You're in a sports car and you don't know how to drive much less what's track and what's field.
You can spend a lot of effort getting good at prompting towards writing tests and E2E tests to at least verify your app does what you expect it to, regardless of experience.
This is a great point and I agree. My own productivity varies based on what part of the codebase I'm working on. If it's "been in there before" and I know the right questions to ask, I can one-shot a good design/improvement. If I'm spending 20-30 minutes asking Fable to "draw a diagram so I can understand" - probably less so. But notably, I CAN get there in a fraction of the time it would have taken before. You can general personalized onboarding docs to ~anything.
Keep the decision-making and execution separate. Use the high IQ models to chat about the design and make them drive subagents to do the actual work. "Chat" style threads are actually quite cheap. Where it gets expensive is having Fable 5 output thousands of lines of implementation where 95% of it was already overdetermined and there were only a few important judgement calls.
I actually have no doubt that I could replace my Opus 5 Low/Medium subagent profiles with Grok 4.5/GLM 5.2/Deepseek v4 Flash and perf would probably be pretty similar.
On top of that - highly recommend adding accurate cost counters to your statusline. You can't improve what you don't measure! (Or even have any intuition about).
It's a fair point, it's not truly unlimited and I do wonder how that would change my workflow. I can definitely imagine if I was inside Anthropic or OAI with unlimited "fast" tokens, you would be more tempted to hand over even more of this process. I completely understand why they talk about "graph engineering" and such, my entire workflow above could be a graph and I could try to increase my leverage even further. Realistically though I am bounded by product decision making, not code output right now.
It’s crazy how we are like ~2y in this AI revolution and still do not have an answer to this question: can you show us the ROI? Where is the revolutionary software your team of agents created?
I found an interesting project recently. As I was looking through the source something felt off. Turned out to be entirely LLM written. There was duplicated code everywhere, same function defined in dozens of files (same name, same intended behavior) but none of them would produce the same output for an input. Dead code all over the place. Over architected. Useless comments.
It was all generated in the last 4 months, so don’t come at me with the “but did they use a model from the last 6 months” nonsense.
There are a surprising number of articles like this along the lines of, "we started using AI tools and ended up spending millions per year".
On what planet do people start paying for things without keeping an eye on the costs and no-one notices until you have spent a crazy amount? I don't understand. You are either paying a fixed amount which you are happy about in-advance or you are PAYG in which case you would ballpark how much it costs.
Otherwise it reads a bit like a fake problem, because it didn't really happen, you just foresaw it (as you should) and added a few guide rails.
There really has never been another product priced like AI is being priced right now. Each of these things has been done before, but all of them together is new.
1. Insanely discounted starter plans. Claude $200/mo plan is like $5k-$8k of API rate usage.
2. Very limited cost visibility, they make it hard to figure out where you spent money (unless you're on the enterprise plan which is for people with unlimited money).
3. Nobody, not even the model provider, knows what your request will cost before it returns. You're writing a blank check every time you hit enter.
4. When you run out you run out very suddenly and disruptively. It's very hard to tell a developer on the 28th of the month "sorry, code by hand until the 1st of next month" so you tend to grant exceptions.
5. The price is changing all the time. New models come in, old models come out, prices change, caching behavior changes, harnesses change, etc. The cost of doing a single task is not predictable even if the task does not change.
6. Basically no volume discounting. Anthropic offered us 2% off for committing to $1M+ per year at API rates.
I manage AI spend for my team at work and I try really hard to keep costs under control but it's absolutely herding cats. Much harder than any other spending I've ever had to manage at work.
Something underlying a lot of this is that pricing models for enterprise coding tools have changed from seat-based to consumption-based pretty quickly, as AI usage has exploded. For months, engineers were able to use unlimited AI for no marginal cost, but that's changed quickly.
In addition, we're seeing people applying AI to more and more use cases, so token growth is very significant. Paired with consumption pricing, it's brought this problem to the forefront very quickly for lots of companies.
The issue is the growth rates can cause costs to drastically change quickly. If you have 1,000 employees and the average is spending $100/month you're at a $1.2M run rate. But suddenly a new model comes out that's twice as expensive, there are some changes to the harness (we found randomly Claude Code and other harnesses will make changes that drastically impact efficiency), and then maybe you have some organic user growth as well and BOOM suddenly you're at a $10M run rate within 60 days. And it's now impossible to forecast future growth.
It is true that this problem can be mostly managed by the techniques we mention here. Those are actually pretty difficult to set up at scale, so many companies (including us) we only really did this in earnest once we started to see those large cost oscillations.
The main reason we shared this here is to maybe help other companies get infrastructure in place before massive cost swings rather than after.
What didn't happen is any analysis of cost/benefit up front. Many of the corporate decisions around AI have seemed characterized by companies blindly copying each other.
Weirdly a lot of the come from company that sell Ai credits in some capacity, and who are also selling (or will soon) some kind of AI gateway or router
Not only this, but perhaps even more nefarious is that AWS gives lots of startups $100k+ in credits. This feels generous when you get it. In reality, it means that (unless you are in a compute intensive startup) you can go for months or years before you hit this, but by the time you do, you already have very solid monthly spend.
Initially, you picked the Multi-ZA RDS db.t3.2xlarge instance because you figured "eh i have credits anyway". Two years later, someone looks at this and says "hey, this is expensive and I bet we can do everything we need on a machine half the size". But then they think "if i downsize it and that works, i'll get a thumbs up emoji on a slack thread. If i downsize it and it causes problems, i'll draw the ire of the whole team. I better leave it alone." And the truth is... by the time your company hits the end of those credits, you're probably at the point where that savings isn't gonna do much. Or maybe you are out of business.
And that is how almost every successful company that uses AWS eventually ends up paying six-figures or more annually.
I suspect that when it comes to hard complex software products, you’re better off ignoring agents and doing “trad coding”. What you lose in short term speed you gain in manageable complex codebases.
If you have a 500k line codebase and even > 50% is written by agents, you are in a world of pain that won’t justify the costs longer term.
Now of course, there are products that just involve lots of code but are not actually complex. This is generally the project with like hundreds or thousands of features but most of the features are separate and don’t actually interact in complex ways. Think a task management app with hundreds of features like calendar, email integration etc. there I think agents gives you more bang for the buck. Just my thought, using agents at work.
Why would it matter if foreign companies analyzed DoorDash data? Pizza deliveries to the Pentagon is all I can come up with, but that's publicly available at https://www.pizzint.watch/
I would bet my entire Polymarket balance ($0) that some military contractors have already asked AIs on the public Internet to design software for them.
1. Codex, Claude and others try to switch models being used at their level itself to manage the cost and outcomes
2. Now company like data bricks develops one more layer on the top of it to do the same task, of finding the base harness and applicable model
Companies like Codex and Claude are focussing/investing heavily on to ensure that people are using their harness directly or instead use APIs. Unless Databricks has some agreement in place they are violating the TOS and openly publishing an article about it. Would be interesting if openAi or Anthropic come back and claim for the API usage prices and all the savings go away.
... where in the article did they say they were using subscriptions? I'm fairly certain enterprises can't access subscription pricing in any case, they're all API costs (Anthropic doesn't support more than 150 on subscription pricing [0][1]).
Going through their harness (codex, claude) is subscription (app use) which is heavily? subsidized.
Anyone using the enterprise plan are charged the API pricing, however the article is not clear if Databricks is using enterprise plan or not which is why added the following disclaimer
Databricks is most certainly getting charged API pricing no matter what harness they are using. OpenAI and Anthropic models are so sought after right now that they set the terms even at the world's biggest companies, there is not a chance to get a special agreement for subscription pricing.
You can use both of those harnesses without going through subscription. That is a native feature in both Codex & Claude Code, even for non-enterprise customers.
This approach seems fundamentally predicated on being able to evaluate coding agents on your own code by having domain specific evals. With that knowledge, you can trust the routing logic is actually improving/maintaining perf while reducing costs.
Without the insight into agent performance, any changes like this feel like a gamble to save $$ at the cost of developer productivity
I'm actually working on building generic repo-specific benchmarks at https://stet.sh ;)
The difficulty of evaluating coding agents is indeed a really big challenge. We built evals on our own codebase and shared some information about that to allow other companies to replicate. We found our own evals correlated loosely with public generic SWE benchmarks.
In large user populations like at Databricks I think the ultimate answer will come from experimentation instead of offline evals. We are already doing this in small groups, exposing them to new candidate models and then measuring per-developer cost and perceived quality changes.
What I take from this is that models are already commoditized, and it’s pretty clear nobody has a moat: routing for the models, they can be swapped whenever new models are released, AI labs will have to continue to run on the treadmill non stop or be replaced. Long term I cannot imagine that business will be high margin. Routing for the harness, so anything that differentiate a provider vs another isn’t exposed to the user and isn’t too relevant.
One more datapoint for the thesis that OpenAI and anthropic aren’t viable, sustainable businesses, and cannot justify their $1T valuation and the level of compute commitment (reminder that OpenAI committed to >$750B in infra spending for 2030)
Do you think Anthropic or OpenAI will eventually try to crack down on routing harnasses? Provide a more vertically integrated experience? They are already trying ro ship hardware products.
I will believe there is no moat when the revenues for Anthropic is not 70B. It seems like people want to throw away money and they don’t like switching
Kudos to databricks, I also find it interesting that such different companies (Stripe, Ramp, Databricks) are all building the exact same internal tools.
I think building companies is going to look more generic in the future because intelligence is an API now.
Thank you for the feedback. We wrote this because after discussing with some of our peer companies, I realized everyone was roughly doing similar things. And I thought it would be good for someone to just systematically write down what those are so that others can try out the techniques if they find them useful.
> nearly every company deploying AI tools at scale has hit the same wall: exponentially growing costs. That curve is unsustainable - left unchecked it will eventually overtake revenue.
But the question remains, AI hasn't shown any explosion in growth/revenue for most companies. The effects seem muted.
It’s funny how different everyone’s experience is with this stuff. To me the diminishing returns are more around not going crazy with prototyping or running with xmax thinking all the time. I haven’t found it hard to stay under the usage limit of one $200/mo Claude and one $200/mo Codex subscription.
If my company told me yeah we’ve decided you don’t get Fable or Opus 5 because it’s too pricey, you gotta use GLM whatever, I’d be displeased.
I feel like I have an advantage over big companies, if I can use the best models on a subscription and not worry about costs much, when they can't do the same as outlined in the article.
> Rapidly adopting newer, more efficient models delivers the largest cost wins of any technique.
I think the more interesting lever is the fourth they mention: token efficiency.
> By the time costly LLM inference occurs, the user's initial statement accounts for only a negligible fraction of the data fed into the AI system, meaning costs are dominated by context the user did not explicitly include.
I think there’s still lots of low hanging fruit in regards to monitoring and improving agent work. Look at your sessions. Look at how much time and context is being spent on, say, a web search returning dozens of results when one good single-pager doc would’ve been better.
100% - there's a lot to learn from traces from real-life sessions with coding tools! For example, I found it pretty eye-opening to see how wide the distribution of tasks truly is. There's also subtle things like how a poorly designed MCP API surface can cause a massive amount of token waste from the model just iterating on finding the right way to call it.
Similar idea re using the same UI across all models, but the agent can modify the harness config as well as start/schedule sessions, etc. The Kanban board feature can be used to orchestrate agent driven workflows, and the agent itself can modify the Kanban lanes. Basically the agent can do all the same things you use the UI for.
Example prompt: “schedule a session using Opus 5 with max thinking for every feature in the PRD on the canvas. For every session make the prompt instruct the agent to review end to end test coverage for the feature and create a report on the canvas for test gaps that you find. Schedule the sessions overnight tonight spread evenly throughout the night, and have each session set to retry upon token exhaustion.”
I think it’s more meta than other meta-harnesses, but I’m biased because it’s my pet project.
I've tested Omnigent superficially, attracted to its thinking around policy, governance, sandboxing, and ui. But it's still alpha at present. I forked its Polly model and got working a somewhat more complex multiagent workflow that I've also modeled in Sandcastle and Gas City but the agent broke after the next update which I would have needed to patch to maintain functionality. Subjectively I also noticed individual models seemed to be performing somewhat worse when wrapped in the platform's framework, presumably due to the extra context introduced (token use was measurably higher). Promising project that I'll revisit when it's further along and I do not doubt the outcomes Databricks claims in committedly dogfooding it.
Double-harnessing distortions seems like a weak point. I wonder if it’s just a temporary measure and long term it’s about writing custom harness going straight to inference APIs across all models.
I've been using it for a week or so. The main draw for me is that I can keep my sessions in one database regardless of the model/provider I use. The webapp can access everything remotely, which is convenient when I'm on my phone.
I haven't gotten a chance to test the multi-agent capabilities, but the DeepSeek Flash prices are so low that I probably will soon.
These seem like the obvious tweaks akin to "using a cheaper hosting platform". I think the real savings come from careful context control for programmatic agents, careful tool awareness and usage to reduce thrashing, distilling workflows into deterministic processes and, moat importantly, adding friction and boundaries for non-technical users who tend to burn tokens making insane asks like "analyze all documents and give me a summary".
How can Smart Router achieve higher task completion rate compared to any of the base models if all it does is dynamically switch base models based on cost??
not sure about the use of exponential and efficiency frontier here, these have formal sides to them but seem to be used rather inflationary and colloquially.
We do this for a lot of our customers (fine tuned to save cost when inference volume is high). Right now for internal coding we are using off-the-shelf models but we are considering fine tuning as well to squeeze more efficiency out.
This has been my experience as well. The best model I have access to right now is Opus 4.8. It's really good at fixing bugs in an established architecture or adding a similar feature, but it's absolutely mid tier at putting together a simple effective architecture for even common software problems. I can tell that functional programming practices are not a large part of it's training material.
Fortunately I’m at a startup with basically infinite Fable 5, and it is legitimately a huge step forward over Opus 4.8. Would recommend. Hopefully some open weight models catch up to its capabilities soon.
how do any of these routing approaches handle kv cache misses? Devin Fusion is the only one that explicitly addresses this, though it does so by switching models during compaction (not sure this isn't still a cache miss though)
We're going to do a followup blog detailing our routing approach soon! In short, the router takes in the task description and infers what models and harnesses are available and makes a recommendation up-front. So essentially the routing decision is made when the harness + model is kicked off and it's only changed halfway through if there's a major delta in complexity from the initial judgment. Therefore, most of the time the cache is maintained just as it would be before (this is the advantage of having a meta-harness that is actually planning all the sub-agents centrally)
Maintaining the cache is extremely, extremely important, so we're iterating fast but that's a major factor we track in the router's development. Couple things I'd look at:
1. The cache is generally reset after a compaction - this is the best time to make a switch if you want.
2. In many cases, the max duration of a cache is 1h, so if a session is being resumed after a long time, that's also a good time to re-assess the complexity.
We're iterating fast here and learning a lot! Definitely a lot to think about it in this area.
The kv cache is wiped as soon as you get your answer, cloud hosts are not going to hold the GPU memory for your entire session. You're probably referring to some agent level cache
Omnigent and OpenRouter are different in the sense that OpenRouter is where you can go to call the actual model but Omnigent is intended to be the place where you go describe the high level task to be done, and work is farmed out to various harnesses and models. Those sandboxes can themselves be using OpenRouter for capacity!
We're calling the layer coordinating harnesses "meta-harness'
Omnigent seems to compete more against Orca https://github.com/stablyai/orca
They both went to be the Agent IDE layer, where you come with your tasks and everything is taken care of. I've been using Orca for a handful of tasks and have been largely enjoying it. My default barebones workflow is ghostty + zmx on ssh connections.
So did we. I just asked my team to get personal accounts that I reimburse them for. It’s just a golden age loop though, the gravy train can’t go on forever unless we start building out thousands of data centers and associated renewable energy.
Really? Because removing it from my company has saved us over 2 million a year and we were able to speed up processing. The chargeback model for databricks is predatory at best.
I think you’ve misunderstood the article. It’s about how Databricks reduced their own costs, not about how adopting Databricks will reduce anyone else’s costs.
I would be really curious to hear from devs at Databricks what the experience of development is like internally. I work at a small startup with essentially unlimited AI spend budget - the entire point is that I should be turning to it at every opportunity since our human labor is so expensive relative to tokens. So generally it's like:
- Spend most time prioritizing/discussing what to do.
- Once that's agreed, use Fable 5 High + 5.6 Sol XHigh come up with a design + plan. Agree on the high level plan. (Usually this just comes down to choosing where the change belongs on the spectrum between minimal patch <-> full redesign)
- Use Opus 5 or Sol Med to execute
- Auto-fix bugs and CI until green + thermonuclear review skill x3.
- Manual interrogation of change/nits
- Come up with QA plan and have Codex Computer Use execute on it
- Manually spot check the final result (usually a sizable diff, thousands of lines, complete feature E2E, etc)
I probably spend like $80 a day at least but I produce the output of 3 or 4 2022 engineers and probably at better quality. So it's easily worth it. Would I save money by switching to GLM 5.2 and such...perhaps? IDK. At our scale it's not worth the time spent building the eval harness to actually understand the performance tradeoff.
I'm also at a startup. My workflow is similar but I have Fable 5 xhigh drive the whole thing: it gets Codex CLI installed in its environment with an API key, and it's instructed to delegate ~everything to Codex and review its work, especially for code quality/conciseness. Fable delegates to Sol or Luna (fast mode) xhigh/max depending on the task - I think Luna xhigh on fast mode is basically a Pareto improvement over Sol medium.
In our team's experience, the product of agents is generally The Homer (1). It does work, but it's vastly overengineered.
When I personally want tight code, I have to spend a considerable amount of time adjusting it manually:
- It needs to be trimmed down. In my experience, at least one agent I use struggles to produce minimalist designs, and it's very frustrating
- I need to consider whether there are solutions based on higher-level assumptions, that AIs typically miss
- I need to check whether there are off-the-shelf solutions - AIs like to reinvent the wheel
IMO, software production has become a mass-produced commodity in every sense - it's much more expensive to produce software manually, but the quality is not the same.
(1) https://simpsons.fandom.com/wiki/The_Homer
It took me a few seconds of deliberating if The Homer was a reference to baseball or "The Odyssey" and then realized there was a footnote
> software production has become a mass-produced commodity in every sense - it's much more expensive to produce software manually, but the quality is not the same.
Agreed.
I believe “The Homer” is a reference to Season 2, Episode 15 of The Simpsons, “Oh Brother, Where Art Thou?”
Homer reunites with his long lost brother, who runs car company Powell Motors. Homer is ultimately tasked by his brother with helping to design a car for the “average man” that ultimately bankrupts the company for being wildly overengineered and costing too much ($82,000 in 1991-money).
Here’s the car: https://simpsons.fandom.com/wiki/The_Homer
If software has become a mass-produced commodity then seems to me the software business will become a much more finance focused business
You will really have to weigh the cost of making the software against the expected revenue.
Since software is still a winner-takes-all market, the mass-production property of software doesn't really matter.
In such markets, what you produce is either worth nothing or worth millions of dollars. For as long as it's the case that well-constructed code (with or without LLM help) is more likely to be in the latter category, the economics of software don't really change.
Even before LLMs, you could've commissioned a half-assed clone of any app you wanted from a 3rd world consultancy for a few thousand dollars. LLMs are basically Bangalore-as-API.
I think part of the reason software was winner take most was the difficulty of making software.
I remember hearing a story that in the past movies were so technically difficult to make that any movie that got made had a good chance to be a profitable hit. But as movies got cheaper to make, more movies got made. Nowadays movie studio execs have to really calculate out the audience and expected revenue for any new movie and balance that against the budget and the cost of the studio's failed movies.
I think a similar dynamic may happen in software
> I think part of the reason software was winner take most was the difficulty of making software.
That might be part of it, but I think it also has to do with the reality of replicating and scaling. Hardware or physical goods simply don’t scale like digital goods. There can be hundreds of knock-off physical products that have lower quality and lower cost but serve 90% of the same purpose, because physical capacity for raw materials, construction, labor, shipping, etc. have scaling limits in each market and economy. Digital goods are just so much easier to replicate and scale, so it often doesn’t make sense to buy software at lower quality and lower price if it doesn’t do most of the job. There are still limits of course, and different from physical goods, but I think this is a key reason why software is seen as winner-take-all.
That honestly sounds a lot healthier than “just ship what the CTO/Product team wants” with as much hand waving as is necessary to very roughly estimate ROI and then pray it hits with the market. In anything that’s not a startup operating in a new industry, the “old way” is a hard way to run a business
> Since software is still a winner-takes-all market, the mass-production property of software doesn't really matter.
Fairly sure software dev was always an iceberg. Most software and most software devs aren't working on horizontal software, but on vertical software, in cost centers. Sales for that kind of software don't scale as much.
Most vertical software is just horizontal software glued together
No surprise, LLM companies optimize for waste. More tokens, and more prompts means more revenue. Reminds of Google’s Prabhakar Raghavan story: deliberately making search worse [1]
[1]: https://pluralistic.net/2024/04/24/naming-names/#prabhakar-r...
Or, more likely, it's that concise code requires a much deeper, wholistic, understanding that these models just are capable of yet.
Same with a junior dev. They don't write long form spaghetti because they're trying to write more LOC. They do it because not doing it is hard, literally above their pay grade.
I use LLM every day, but they're still completely awful at architecture. I don't think this clear lack of ability is some conspiracy.
As a business user, the same thing is true for non-code documents. The biggest exertion is reducing the excessive slop down to concise, clear points.
> software production has become a mass-produced commodity
For who?
The public? The public has never liked buying software at any price.
Businesses? Businesses need higher quality software when it's relevant to their core competencies, so they hire people instead. Buying competing SaaS or depending too much on AI is throwing the baby out with the bathwater.
This was more true a few months ago but Fable has improved the situation considerably.
Also just remember - minimalist code looks and feels great but customers do not read your code. I have caught myself many times providing "corrections" to abstractions that were already ~fine, just not perfect. The average SWE costs $200/hr. Careful you don't burn $50 worrying about code that will likely be rewritten or can be better abstracted when that's actually needed.
> The average SWE costs $200/hr
This is a pointless quibble but the hourly rate claim is not true--it's like ~$60 in the USA [0]. Maybe you meant at a specific Org but this is important context when comparing "pricing" between human and AI.
[0] https://www.salaryexpert.com/salary/job/software-developer/u...
How is this not true? Taking a Senior SWE @ ~$200K, even just the base salary cost / 2080 working hours is $100/hr. Fully loaded employer cost + accounting for non-coding time gets you to upper 100s easily.
Even for a junior making $100K, I have a hard time believe their time is worth less than $75/hr or so.
Edit: Fine, "Senior" is not "Average". But naive salary is not the true numerator.
Western Europe is mostly consultancy, and the rate paid by client is usually higher, and doesn't matter if it's eastern Europe, Portugal or even India.
It is. And the quality is on par with any us eng. People here forget that the big comp packages are a minority even in the US. The cost tho is much higher than just salary.
Company time != Pay rate, if you're working somewhere that's publicly traded check out "revenue per employee" metrics sometime.
I hire contractors for a large enterprise in the US. The going rate is typically $85-$100/hr for a senior dev, depending on specialization. Lead-level maybe $120 for the right skill set.
Of course, the SWEs making that much (over 200k) are not representative of the broader field. That's the point.
Pay hits a ceiling, and that ceiling is moving lower regardless of experience. That has nothing to do with AI, but what the market will bear. Hiring counts of humans must increase no matter what. Moving some of the spend to AI reduces the risk of hiring less qualified employees they might have rejected a decade ago.
Wages at the top end are stagnating to subsidize this. That's undeniable.
$15, where we're going.
Federal minimum wage is $7.25 per hour
An MBA's rule of thumb is that a full time employee's hourly cost to a business is at least 1.5x to 2x times their salary depending on employer taxes, benefits, offices, travel, training, hardware, perks, etc.
Minimalist code is necessary to keep AI agents working well for longer than a month on a system IME. At a certain point, their own machinations overwhelm them and they both slow down, and make worse and worse decisions.
Also, you can probably think about it like this:
"Will I benefit from this code being minimalist before [date]", where [date] is whenever you think the agent will be good enough to come back and make the corrections you would make today.
I'd caution that some corrections become harder to make over time, rather than easier. A bad architecture now can become much harder to fix once other things have grown up around it.
Even as a self-contained unit, you can't step in the same river twice, and on [date] some important details may have seriously faded, both in terms of text that can be mined and also in terms of human "why did we do that" and "what was the reason we did it this way and not that way" etc.
> The average SWE costs $200/hr.
And this is how I find out I'm woefully underpaid.
That "cost" includes all the overhead provided by the company: benefits, rent for offices, utilities, equipment, etc. The average SWE is not taking home anything close to that, outside of Silicon Valley and a few other limited areas.
Whatever you are making this year as SWE you'll be making less next year if the current trend in improvement of AI coding aids is going to be sustained. Think about it : programmers used to derive a lot of their value from the fact that it was a hard skill to acquire. My kids can now 'vibe code' stuff faster (and better looking) than what I could come up with as the beginnings of a design plan. And then I still need to implement it.
There's a massive difference between your kids vibe coding something and an engineer using AI to implement something. If you're unable to discern the difference, that's something to reflect on :)
It doesn't matter if GP is able to discern the difference, it matters if your CEO is forced to care about the difference.
CEOs will always want someone who knows to implement so we're safe from kids vibe coding their way in but in a short while it becomes AI who knows who is managing less expensive AI.
The average SWE makes $400k a year? Are you being serious?
Costs, not makes. That includes employer taxes, benefits, offices, travel, training, hardware, any perks.
IME this works until it does not. This approach works well at the beginning of a greenfield project, but at the same time because it is so easy to add features, you will likely ship something that is way too over engineered. And that complexity will not amortize over next increments and will more likely lead to the entire project being a black box only fully understood by AI. However a more careful use of AI for targeted surgical changes is far more ”productive” in the long term IMO.
I disagree, the approach works well in a legacy project, since there are structures and standards that already exist, which them model can draw from (if you aren't more explicit about it in AGENTS.md)
Disagree. I operate this way inside a multi-million line legacy codebase.
> I work at a small startup
How does a “small startup” end up with a multi million line “legacy” codebase? Something not mathing
> How does a “small startup” end up with a multi million line “legacy” codebase?
Easy! The output of 6 months ago Opus! Which seemed so wonderful at the time.
This! I don't think folks understand how easy it is to go from greenfield to brownfield with these tools, esp if your organization is only valuing velocity. Meaning your doing full agentic development on large features, barely reviewing any code, and shipping without much refinement. It's insane, but this appears to be the status quo in SF startups.
Have you worked at many startups?
Something isn't clear about the size of your codebase here and the level of reliability your customers expect, as a reader of your comments. Clarity there will help.
My observation has been:
- Initial greenfield work by an LLM is fast and very effective with minimal or no human oversight.
- Subsequent work ends up being over engineered and very verbose. Assumptions are made that aren't suited to the problem at hand (for example I find Fable is extremely regex happy where structured data would work much better from a readability perspective.)
- Once code bloats beyond a certain point due to unguided LLM usage, complexity is high enough that only LLMs can operate on the codebase with any economical amount of time.
- Rinse repeat and your code ends up unclear about any state that's not explicitly being tested and verified in QA loops
For some of our products this has been fine, for others it's been problematic. An understanding of your size and reliability requirements will help make the conversation more productive.
> unguided LLM usage
Why aren't you guiding your LLM usage? Is that what I said - to spam it and not guide anything? Or to have a careful workflow where you agree on design and maximize your human judgement/leverage?
> any state that's not explicitly being tested and verified in QA loops
As opposed to before, when engineers perfectly reasoned about code behavior from first principals and QA was unnecessary?
> Why aren't you guiding your LLM usage? Is that what I said - to spam it and not guide anything? Or to have a careful workflow where you agree on design and maximize your human judgement/leverage?
You didn't say anything positively or negatively regarding this so I made an assumption that you were using the LLM relatively unguided (e.g. a bit of oversight, not the kind of thing that heavy code reviews used to involve pre-agents.) Feel free to add clarity on your actual usage loop.
> As opposed to before, when engineers perfectly reasoned about code behavior from first principals and QA was unnecessary?
In my experience, most engineers are quite good at reasoning about code behavior for non-QAed code paths. Obviously things fall through the cracks. But I've been in the ground floor of plenty of Big Techs in their early stages before agents and, yes, a lot of initial development had spotty test coverage and yet most of the engineers had good mental models of what was happening. It used to be a very valuable skill to wrap your head around a torrid piece of code with few or no tests but was nonetheless a core piece of your application. Conversely, agentic development can bring cognitive debt [1].
===
This isn't a fight. We aren't sparring over what's right and wrong. I'm just curious how other people use agents in their work as someone who is also now in a startup that uses LLM agents heavily and has no limitations on spend.
[1]: https://martinfowler.com/fragments/2026-02-09.html
> You didn't say anything positively or negatively regarding this so I made an assumption that you were using the LLM relatively unguided
I feel like this statement betrays your lack of advanced experience coding with LLMs.
OP's elaboration of the steps they are going through (planning, agreeing on plan, getting one LLM to draft execution plan, approving it, then executing with a separate LLM, then reviewing/testing) made it super obvious to me that they are guiding their LLMs quite considerably as part of their work.
Anyone making blanket statements about LLMs producing garbage is just telling on themselves about not having proper SDLC practices in place.
Planning, agreeing on a plan, separating planning and implementation LLM, using separate review LLMs, these are all table stakes. This isn't "guidance" if you're getting paid to write software. If you think "unguided" means "I typed a prompt into claude code and waited yolo" I don't know what to say but, you have a very different idea of what professionals do than I do.
I find for my own work that I need to read the diff the LLM produces then offer feedback on the diff in its own loop before I am satisfied, and this is after all the unattended QA steps through Codex Computer or Claude MCPs happen. Then auto reviewers come in and then reviewers come in. Of course, at our stage, we rarely have this luxury and it's only reserved for the very core of our codebase.
This is still much less guidance than we used to do for code before agents became popular. Even at Series A companies, before agents, we used to socialize tech specs, get buy-in from multiple engineers, create test plans, etc etc.
> Anyone making blanket statements about LLMs producing garbage is just telling on themselves about not having proper SDLC practices in place.
> I feel like this statement betrays your lack of advanced experience coding with LLMs.
Are we in school debate club? I don't know what's going on lol, I'm just curious how people are using LLMs! Is it just that irresistable to take a cheap shot at each other?
> Are we in school debate club?
Not that I know of but that's the conclusion I drew from your statement.
It's not a cheap shot unless you took it personally?
I suppose I could have said "the fact that OP's explanation of how they work did not lead you to conclude they were in fact guiding their LLM usage quite a bit tells me that perhaps you have not been working with LLMs in any advanced capacity".
For the SDLC comment I admit it was a broader statement (based on observing people generalizing that "LLMs produce bad outputs") and not specifically aimed at you, and I didn't make that clear, so my bad.
> "unguided" means "I typed a prompt into claude code and waited yolo"
Yes, this is literally what that means.
> If you think "unguided" means "I typed a prompt into claude code and waited yolo" I don't know what to say but, you have a very different idea of what professionals do than I do.
What exactly does "unguided" mean to you, then?
AI is an accelerate tool for any organizations, management thinks it'll solve their organization issue because it accelerates it. Most often, it accelerates toward a wall.
Design is too expensive, we do agile. QA too expensive, we fire all of them, and claim devops is the now, which allows us to fire the Ops team too, 100% ownership from deisng to ops on devs.
One person with an agent can replace all these teams. Yeah mo profits.
No, but not relevant.
What is the point of working at a startup if you’re dealing with millions of lines of legacy code ? Isn’t the whole point of startups to create & innovate with a clean slate and modern tools?
No, actually. The point is to build a profitable business.
How long has your startup been around? I’ve worked at plenty of startups over the past 20 years. Including one that was still calling themselves a startup 10 years out. The org I work at now was a startup before my tech giant employer acquired them. We have a very bloated and very profitable 8 year old codebase that is barely 500k LOC.
I’ve never seen a startup with a multi million line legacy codebase.
They may have forked something
Definitely possible, but up thread they wrote:
>”Have you worked at many startups?”
In response to a question about a legacy codebase at a startup. That implies that they think whatever they are doing is common. And forking a multi million line codebase and heavily developing it isn’t common for startups.
I don't know if that's the whole point, but I agree with the sentiment, why would a startup be working in legacy code and where would that code come from if this is truly the start of something.
OP might just be working at a small software company or for one that broke from a bigger one and is now "startup" like?
You'd be surprised. I met a guy last week who was proud to tell me he had vibe coded an almost 2 million line code base. The app did not sound that complicated, so I'm assuming it's full of copy-pasta flavored slop.
"startup" and "legacy codebase" are diametrically opposed concepts.
And if you're saying (based on your other comments) that a 6 month window is enough to create a legacy codebase...that indicates a serious lack of experience or understanding as to what a legacy codebase is, or why they exist.
Man, so many people in this thread just arguing pointless semantics, making weirdo absolutist (and incorrect) statements.
Accept that other people may ascribe different meanings/interpretations to words than you, and that if your reading of their statement doesn't make sense to you, perhaps you are simply reading it wrong.
Trying to hold someone else to your definition of words suits what purpose exactly? Are you just trying to "win" ?
Yes lol. Of all things people are getting on me for it's the number of LoC x Years In Business of this startup. I don't fucking know, I didn't start the company and I wasn't here for several of those industrious years. Looking now it looks like we have slightly fewer LoC than that, I was counting some of the generated stuff.
But who cares? The point is any codebase over a few years old with lots of customers and a big surface area has lots of code, much of it "legacy" from the standpoint of a guy in 2026.
You are wrong and they are wrong but that is ok.
This is an ultra cop out. There are standards in language that are not all “left means right for me so you cannot assume when i say right it is right and not left”
This whole thread around loc is depressing. It speaks volumes of some peoples inexperience working on actual legacy code. Legacy code is not just age or size but that the technical foundation is dated in a fundamental way. A giant monolith running on a now defunk framework using a database only one guy in canada knows about.
Case and point in my day job. The org that owns XMM development does not know how to recover a physical bench that is bricked because everyone who knew how has left. So now they just use simulators…
Interestingly, AI figured out some of this pretty easily for me. But the org has the exact same AI as i do. At the same time another org is close to a year into a greenfield rewrite that has been developed via agentic swarms. Absolute trainwreck.
AI doesnt make bad engineers good. Anyone who says they are doing 4 eng work likely would be without ai too. Those that claim otherwise, are the bad engineers.
I agree with your larger point. Though I think it's pretty natural to wonder how the poster ended up with a multi million line codebase.
2M lines of code is 15 people committing ~26k lines of code per year (~100 lines per working day) for 5 years.
15 people is a pretty small startup, what if this is a 50-person startup?
Doesn't seem like that much to me, depending on what you're building and the size of your team.
exactly. Usually legacy code forms when people lose context and confidence in parts of the codebase due to staff turnover etc and ppl avoid touching or enhancing those parts for long periods. Six months is a short time to accrue that much tech debt, its enough time where most of the people who created that "legacy" are probably still around. As you said indicates bigger problems.
So basically any LLM codebase of sufficient size is immediately legacy.
Wow how many years of experience with Claude Code and Codex do you have? lol
The job requires 10 years of those technologies ;)
I'm probably between $50-$200/day depending on the day; we also have effectively unlimited budget, though a lot of that is because Azure gives startups $150,000 in credits for 2 years, which we've wired up to a LiteLLM gateway & OpenCode. Without that I think our appetite would be more around $400/month/employee.
A lot of my high costs is because I just throw Sol at everything. If I were more selective and brought in Luna or v4 Flash every once in a while, I think I'd be more like ~$400/month. That's why I'm not aligned with the notion that "tokens are subsidized so that's why people are using so much": its not that I'll have to adjust to using less, its just that I'd need to think before I prompt a bit and be more judicious. I could easily see my raw token counts doubling or tripling in the coming months. I don't think that will change as subsidization subsides; though maybe lab revenue will; intelligence per dollar is getting cheaper every week. Its solely a function of adaptation to process, which takes time.
The productivity gains per token are the single most asymmetrical thing I've ever seen in engineering. The engineers on our team are pretty effective with tokens; easily that 2x-4x output as you're seeing, spending $20-$200/day. Some of our security folks have also started contributing more-and-more code, and they're on the other side: they'll spend hundreds a day running in circles, eventually producing these +/-30k loc pull requests that take ages to get merged and are littered with issues. They weren't writing much code before, so arguably they're more productive by some multiplier greater than 1, but I think the drag on the rest of the team, and potential issues with what they produce, has overall created a net-negative situation. Inversely, some other company functions have produced a few one-off websites for things like sales processes, and those have been a huge win. The asymmetry is wild. There's almost a valley of incoming skill where if you know nothing about code, you'll leverage it well; if you know just a little bit, it makes you super dangerous; if you know a lot, you're the biggest winner. Really difficult situation to navigate.
This is very close to my workflow but you forgot one important step:
- Suggest a better approach that makes the AI say, “That’s much simpler. And you’re right. My original plan was over-engineered.”
Do you have issues with performance at the moment? Right now I tend to find that it produces absolutely terrible design patterns and especially performance. I mean maybe I don't know exactly what area you're looking at but yeah for us we tend to find it's terrible wrt dB/caching/scaling and often any performance improvements it proposes end up actually shooting itself in the foot and being worse than before but it's not very good at testing in an organized way to even notice it made it worse despite repeated prompts to do so I mean if I prompt it to test performance in a handheld structured way (it is very bad at finding out what performance to test and why) before making changes I can usually figure it out but it usually takes insistence on the specifics to really ensure a good solution that will actually fix the problem
Performance is better than ever. It's never been more practical to set up wildly complex synthetic test environments and measure perf wins. Plus the models will find every possible algorithmic/design improvement.
It actually gives me quite an uncanny feeling, bulldozing over years of human optimization work with a newer, "perfect" design. Like bringing an AK-47 back to the middle ages.
> the models will find every possible algorithmic/design improvement
it's so hard to square such totalizing statements with my day to day experience with fable and sol, (every possible, improvement, really?? they are NOT omniscient) arguing with them/my colleagues' agents that no they have slowed down the system 200x with their terrible change, doing string operations on millions of db rows, trying to get it to understand that I don't care that it's calling it a "cache" if a cache hit is slower than what we had before.
These agents do let you learn codebases quickly, and produce code way faster. I don't look at IDEs all that often. But literally multiple times every single day I catch them doing something stupid.
I don't think its impossible that we could get better performance from the agents. I know ive tried all sorts of workflows and skills, few of which seem to have much effect on the things the models struggle with. I think a big part of it is encoding enough context for large codebases, and providing it with all the tools it needs to make it successful, things to automatically check its work, etc. But that's not automatic, in fact its generally a terrible judge of what it needs or what its bad at
yeah it's true, you do have to guide them. i find that the key is you have to know what's possible. you have to have the instinct for "this really shouldn't be so difficult". my junior SWE coworkers have the same trouble as your coworkers.
but the revolution is it doesn't take that long. in like 15 minutes you can chat with fable and get to the meat of whatever the issue is with repeated questioning. and then it does the solution for you. so it's not magic but it's still like a 100x speedup.
I needed to thoroughly test rerankers on my companies rather unique corpus.
Opus and I wrote a parallelized test harness and labeled groundtruth in around 2 hours.
In 2022 that would've likely been all I did for a couple sprints
I encounter this regularly and it still feels weird.
That sense that you did something better in a few days than you would have in a month 5 years ago. It's like buying a table saw for wood working.
One crazy thing I think about often is how there are so many correctness and testing harnesses that would have taken weeks to build in the past so we simply never would have. We'd just do our best then wait and see what comes to the surface. This is a huge part of what makes it possible to actually make better software with LLMs in my opinion. It isn't just 'LLM codes better than I ever could' (that's often untrue still) but 'LLM enables me to make assertions about the program to degrees that would have been absurdly impractical in the past'. It's huge
Yes 100%. This morning I casually prompted Codex to drive the browser to complete extensive performance testing in-situ that would have literally been weeks of work before. Probably in reality it just wouldn't have been done, and performance guarantees would have been attempted up front via more careful design.
In this case the design was also AI generated, and there were limited wins to be found because the design was already superb.
Are you using the SOTA models at very high reasoning during planning? IME that makes a LOT of a difference. I‘d also never let them just rip into the architecture, but always push back and ask for alternatives first. Once the overall plan is nailed, not that much can go wrong. Provided it’s a reasonable change set and not a 20k LOC PR.
fable or sol w/ very high both planning and execution, yeah. I feel the "push back" part is a big part of my job now (on every step, planning, execution, and review) yeah, but that feels pretty incompatible with the sorts of "just let it do what it wants" which other people seem to be claiming
You have an unlimited budget, and you only spend $80/day? I’m up to $3k/week, and still expanding.
> Spend most time prioritizing/discussing what to do.
you should probably be doing this discussion work along with Fable 5. It will give good feedback if you're working on the correct things.
> Come up with QA plan and have Codex Computer Use execute on it
QA plan should be part of the above "design + plan", not after it. The implementer needs to be able to fully test before publishing a PR. This is true whether humans or agents are writing the code.
> Manually spot check the final result (usually a sizable diff, thousands of lines, complete feature E2E, etc)
unfortunately this is not really scalable with amount of code agents can produce, so you need independent (fresh context) agent reviewers to help. Ideally they only escalate to a human when really stuck.
> I probably spend like $80 a day at least
at a small startup you should be on the $200/month plan(s).
$80 sounds extremely low for what you're describing - are you on API token plans?
I have had some $3,000 token days - even without Fable. I don't see how this is sustainable.
My personal 20x plans get so much usage for so cheap. The consumer subsidies are crazy, but alas I can't use them for work.
$80 is definitely low now that I look at my numbers. but not OOMs low, it's closer to like $200 on heavy days. i don't know how you're doing $3k/day, that's wild. i'm pretty aggressive about compaction and session restarts, and i reserve Fable 5/Sol XHigh for "main thread" orchestration
Output of 3 or 4 2022 engineers? Its that your self assessment? Output as in number of lines of code?
Only spending $80 a day on Opus 5/Fable 5/GPT 5.6 Sol feels very low. I'll roll through a couple hundred dollars worth of credits a day with those models, the vast majority of which would be on non-coding tasks, and it's still a huge cost savings over me or my team having to do these things manually, if we'd even be able to do them at all.
But that's also why it's now easy to justify the cost of an Nvidia or Intel inference server with Kimi K3 locked and loaded :)
I was just about to say, how could routing possibly be worth it at the risk that the work output is sub par?
The output yes, but do you produce the impact and value of 3 engineers? I have seen this workflow being toyed with too, and I find it to produce massively overengineered stuff that actual people don't really wanna use
He only spot checks 1000s loc diffs, so probably has no clue.
My experience is that your description works for a certain time, since you're knowledgeable of the codebase and can guide it. But after too many iterations with not hand-holding the llm, it quickly gets unwieldy.
>Auto-fix bugs and CI until green + thermonuclear review skill x3.
Gotta love this loop, I have it running while I'm asleep all the time.
Do you have tips for generating clean productive output per dollar?
in my humble experience it boils down to mastery.
Are you at least conversational in the subject matter? You're gonna have a good time just by paying attention and adjusting your workflow. If you're getting a lot of back and forth with it, its asking a lot of planning type questions, stop, step back, rethink the whole feature, and start again from the beginning with everything more fleshed out.
If you are in a brand new field, there's no way to bridge that divide. The issue is you don't know what is good or bad, or whether what you have learned is good or bad. You're in a sports car and you don't know how to drive much less what's track and what's field.
You can spend a lot of effort getting good at prompting towards writing tests and E2E tests to at least verify your app does what you expect it to, regardless of experience.
This is a great point and I agree. My own productivity varies based on what part of the codebase I'm working on. If it's "been in there before" and I know the right questions to ask, I can one-shot a good design/improvement. If I'm spending 20-30 minutes asking Fable to "draw a diagram so I can understand" - probably less so. But notably, I CAN get there in a fraction of the time it would have taken before. You can general personalized onboarding docs to ~anything.
I appreciate this non-judgmental description of what it's like to approach a topic/technology from a newcomer's perspective. Thanks!
Keep the decision-making and execution separate. Use the high IQ models to chat about the design and make them drive subagents to do the actual work. "Chat" style threads are actually quite cheap. Where it gets expensive is having Fable 5 output thousands of lines of implementation where 95% of it was already overdetermined and there were only a few important judgement calls.
I actually have no doubt that I could replace my Opus 5 Low/Medium subagent profiles with Grok 4.5/GLM 5.2/Deepseek v4 Flash and perf would probably be pretty similar.
On top of that - highly recommend adding accurate cost counters to your statusline. You can't improve what you don't measure! (Or even have any intuition about).
> essentially unlimited AI spend budget
> I probably spend like $80 a day
This doesn’t sound like “unlimited”, I spend more than this out of pocket per day and I have a strict budget.
It's a fair point, it's not truly unlimited and I do wonder how that would change my workflow. I can definitely imagine if I was inside Anthropic or OAI with unlimited "fast" tokens, you would be more tempted to hand over even more of this process. I completely understand why they talk about "graph engineering" and such, my entire workflow above could be a graph and I could try to increase my leverage even further. Realistically though I am bounded by product decision making, not code output right now.
but I produce the output of 3 or 4 2022 engineers and probably at better quality.
Possibly, but the output of a 2022 engineer is about 1/10th of the output of a 2010 engineer, so it's an extremely low bar.
also - as always with these claims there's no actual product / repo / whatever one could check.
I would love to see what these tools create but outside slop there's never: This works, is in production, here's the code.
Any day now.
It’s crazy how we are like ~2y in this AI revolution and still do not have an answer to this question: can you show us the ROI? Where is the revolutionary software your team of agents created?
I found an interesting project recently. As I was looking through the source something felt off. Turned out to be entirely LLM written. There was duplicated code everywhere, same function defined in dozens of files (same name, same intended behavior) but none of them would produce the same output for an input. Dead code all over the place. Over architected. Useless comments.
It was all generated in the last 4 months, so don’t come at me with the “but did they use a model from the last 6 months” nonsense.
do the same across 20 terminals (as you should) and now you are up to $1.6k/day. would that give you output or 60-80 engineers? not a chance, right?
no one’s AI spent will be in question working a single terminal with carefully planned out and executed process you do
There are a surprising number of articles like this along the lines of, "we started using AI tools and ended up spending millions per year".
On what planet do people start paying for things without keeping an eye on the costs and no-one notices until you have spent a crazy amount? I don't understand. You are either paying a fixed amount which you are happy about in-advance or you are PAYG in which case you would ballpark how much it costs.
Otherwise it reads a bit like a fake problem, because it didn't really happen, you just foresaw it (as you should) and added a few guide rails.
There really has never been another product priced like AI is being priced right now. Each of these things has been done before, but all of them together is new.
1. Insanely discounted starter plans. Claude $200/mo plan is like $5k-$8k of API rate usage.
2. Very limited cost visibility, they make it hard to figure out where you spent money (unless you're on the enterprise plan which is for people with unlimited money).
3. Nobody, not even the model provider, knows what your request will cost before it returns. You're writing a blank check every time you hit enter.
4. When you run out you run out very suddenly and disruptively. It's very hard to tell a developer on the 28th of the month "sorry, code by hand until the 1st of next month" so you tend to grant exceptions.
5. The price is changing all the time. New models come in, old models come out, prices change, caching behavior changes, harnesses change, etc. The cost of doing a single task is not predictable even if the task does not change.
6. Basically no volume discounting. Anthropic offered us 2% off for committing to $1M+ per year at API rates.
I manage AI spend for my team at work and I try really hard to keep costs under control but it's absolutely herding cats. Much harder than any other spending I've ever had to manage at work.
I think my startup can help: https://unbiased.ai
Happy to give you (or anyone here) some trial credits if interested! Email address in my profile.
Something underlying a lot of this is that pricing models for enterprise coding tools have changed from seat-based to consumption-based pretty quickly, as AI usage has exploded. For months, engineers were able to use unlimited AI for no marginal cost, but that's changed quickly.
In addition, we're seeing people applying AI to more and more use cases, so token growth is very significant. Paired with consumption pricing, it's brought this problem to the forefront very quickly for lots of companies.
The issue is the growth rates can cause costs to drastically change quickly. If you have 1,000 employees and the average is spending $100/month you're at a $1.2M run rate. But suddenly a new model comes out that's twice as expensive, there are some changes to the harness (we found randomly Claude Code and other harnesses will make changes that drastically impact efficiency), and then maybe you have some organic user growth as well and BOOM suddenly you're at a $10M run rate within 60 days. And it's now impossible to forecast future growth.
It is true that this problem can be mostly managed by the techniques we mention here. Those are actually pretty difficult to set up at scale, so many companies (including us) we only really did this in earnest once we started to see those large cost oscillations.
The main reason we shared this here is to maybe help other companies get infrastructure in place before massive cost swings rather than after.
Someone did notice, as they panicked at the cost.
What didn't happen is any analysis of cost/benefit up front. Many of the corporate decisions around AI have seemed characterized by companies blindly copying each other.
Weirdly a lot of the come from company that sell Ai credits in some capacity, and who are also selling (or will soon) some kind of AI gateway or router
> we started using AI tools and ended up spending millions per year
This is how AWS made its fortune.
Not only this, but perhaps even more nefarious is that AWS gives lots of startups $100k+ in credits. This feels generous when you get it. In reality, it means that (unless you are in a compute intensive startup) you can go for months or years before you hit this, but by the time you do, you already have very solid monthly spend.
Initially, you picked the Multi-ZA RDS db.t3.2xlarge instance because you figured "eh i have credits anyway". Two years later, someone looks at this and says "hey, this is expensive and I bet we can do everything we need on a machine half the size". But then they think "if i downsize it and that works, i'll get a thumbs up emoji on a slack thread. If i downsize it and it causes problems, i'll draw the ire of the whole team. I better leave it alone." And the truth is... by the time your company hits the end of those credits, you're probably at the point where that savings isn't gonna do much. Or maybe you are out of business.
And that is how almost every successful company that uses AWS eventually ends up paying six-figures or more annually.
On this planet?
They’re not saying they regret doing it, or that it was a mistake.
They’re just saying they’ve gained experience and have leveraged the tools to an extent their usage can be optimized.
Pretty standard business or life iteration.
I suspect that when it comes to hard complex software products, you’re better off ignoring agents and doing “trad coding”. What you lose in short term speed you gain in manageable complex codebases.
If you have a 500k line codebase and even > 50% is written by agents, you are in a world of pain that won’t justify the costs longer term.
Now of course, there are products that just involve lots of code but are not actually complex. This is generally the project with like hundreds or thousands of features but most of the features are separate and don’t actually interact in complex ways. Think a task management app with hundreds of features like calendar, email integration etc. there I think agents gives you more bang for the buck. Just my thought, using agents at work.
> if you have a 500k line codebase and even > 50% is written by agents, you are in a world of pain that won’t justify the costs longer term
A bold claim to make with little to no supporting evidence
Careful. If you admit to using models that weren't trained by OpenAI or Anthropic then you might hauled in front of Congress: https://www.scmp.com/news/china/diplomacy/article/3362616/us...
I'd prefer congress to be asking questions (this is all they are doing so far, based on the article) before doing any legislating.
Why would it matter if foreign companies analyzed DoorDash data? Pizza deliveries to the Pentagon is all I can come up with, but that's publicly available at https://www.pizzint.watch/
I would bet my entire Polymarket balance ($0) that some military contractors have already asked AIs on the public Internet to design software for them.
I find this funny and interesting at some levels
1. Codex, Claude and others try to switch models being used at their level itself to manage the cost and outcomes
2. Now company like data bricks develops one more layer on the top of it to do the same task, of finding the base harness and applicable model
Companies like Codex and Claude are focussing/investing heavily on to ensure that people are using their harness directly or instead use APIs. Unless Databricks has some agreement in place they are violating the TOS and openly publishing an article about it. Would be interesting if openAi or Anthropic come back and claim for the API usage prices and all the savings go away.
... where in the article did they say they were using subscriptions? I'm fairly certain enterprises can't access subscription pricing in any case, they're all API costs (Anthropic doesn't support more than 150 on subscription pricing [0][1]).
[0]: https://support.claude.com/en/articles/9797531-what-is-the-e...
[1]: https://support.claude.com/en/articles/9266767-what-is-the-t...
Going through their harness (codex, claude) is subscription (app use) which is heavily? subsidized.
Anyone using the enterprise plan are charged the API pricing, however the article is not clear if Databricks is using enterprise plan or not which is why added the following disclaimer
> Unless Databricks has some agreement in place
Databricks is most certainly getting charged API pricing no matter what harness they are using. OpenAI and Anthropic models are so sought after right now that they set the terms even at the world's biggest companies, there is not a chance to get a special agreement for subscription pricing.
You can use both of those harnesses without going through subscription. That is a native feature in both Codex & Claude Code, even for non-enterprise customers.
They certainly have an enterprise plan?
Databricks will be using the API anyway, thats all you get with an enterprise agreement.
Open ai is allowing subscription use, anthropic also paused the effort to stop subscription use.
They did? Is there a source where I can learn more? I'd love to use my Anthropic subscription with opencode.
Only if opencode uses the agent sdk/Claude -p
June 15 changes would be the keyword to check, but at least in anthropics case it's... Complicated
Bans are not in effect?
buddy, they're on enterprise plans paying per token
This approach seems fundamentally predicated on being able to evaluate coding agents on your own code by having domain specific evals. With that knowledge, you can trust the routing logic is actually improving/maintaining perf while reducing costs.
Without the insight into agent performance, any changes like this feel like a gamble to save $$ at the cost of developer productivity
I'm actually working on building generic repo-specific benchmarks at https://stet.sh ;)
The difficulty of evaluating coding agents is indeed a really big challenge. We built evals on our own codebase and shared some information about that to allow other companies to replicate. We found our own evals correlated loosely with public generic SWE benchmarks.
In large user populations like at Databricks I think the ultimate answer will come from experimentation instead of offline evals. We are already doing this in small groups, exposing them to new candidate models and then measuring per-developer cost and perceived quality changes.
What I take from this is that models are already commoditized, and it’s pretty clear nobody has a moat: routing for the models, they can be swapped whenever new models are released, AI labs will have to continue to run on the treadmill non stop or be replaced. Long term I cannot imagine that business will be high margin. Routing for the harness, so anything that differentiate a provider vs another isn’t exposed to the user and isn’t too relevant.
One more datapoint for the thesis that OpenAI and anthropic aren’t viable, sustainable businesses, and cannot justify their $1T valuation and the level of compute commitment (reminder that OpenAI committed to >$750B in infra spending for 2030)
Do you think Anthropic or OpenAI will eventually try to crack down on routing harnasses? Provide a more vertically integrated experience? They are already trying ro ship hardware products.
I will believe there is no moat when the revenues for Anthropic is not 70B. It seems like people want to throw away money and they don’t like switching
Surprisingly pragmatic and info packed article..
Kudos to databricks, I also find it interesting that such different companies (Stripe, Ramp, Databricks) are all building the exact same internal tools.
I think building companies is going to look more generic in the future because intelligence is an API now.
Thank you for the feedback. We wrote this because after discussing with some of our peer companies, I realized everyone was roughly doing similar things. And I thought it would be good for someone to just systematically write down what those are so that others can try out the techniques if they find them useful.
+1 well written, well paced article. Pleasure to read.
Have you tried measuring Gemini? now that you have the router it should be a simple task. Thanks!
> nearly every company deploying AI tools at scale has hit the same wall: exponentially growing costs. That curve is unsustainable - left unchecked it will eventually overtake revenue.
But the question remains, AI hasn't shown any explosion in growth/revenue for most companies. The effects seem muted.
It’s funny how different everyone’s experience is with this stuff. To me the diminishing returns are more around not going crazy with prototyping or running with xmax thinking all the time. I haven’t found it hard to stay under the usage limit of one $200/mo Claude and one $200/mo Codex subscription.
If my company told me yeah we’ve decided you don’t get Fable or Opus 5 because it’s too pricey, you gotta use GLM whatever, I’d be displeased.
I authored this - happy to answer any questions.
As a solo dev, this gives me hope.
I feel like I have an advantage over big companies, if I can use the best models on a subscription and not worry about costs much, when they can't do the same as outlined in the article.
> Rapidly adopting newer, more efficient models delivers the largest cost wins of any technique.
I think the more interesting lever is the fourth they mention: token efficiency.
> By the time costly LLM inference occurs, the user's initial statement accounts for only a negligible fraction of the data fed into the AI system, meaning costs are dominated by context the user did not explicitly include.
I think there’s still lots of low hanging fruit in regards to monitoring and improving agent work. Look at your sessions. Look at how much time and context is being spent on, say, a web search returning dozens of results when one good single-pager doc would’ve been better.
100% - there's a lot to learn from traces from real-life sessions with coding tools! For example, I found it pretty eye-opening to see how wide the distribution of tasks truly is. There's also subtle things like how a poorly designed MCP API surface can cause a massive amount of token waste from the model just iterating on finding the right way to call it.
First time hearing of Omnigent. Anyone have experience using it?
If you’re into web based meta-harnesses you might like Circus Chief: https://github.com/ferrislucas/Circus-Chief
Similar idea re using the same UI across all models, but the agent can modify the harness config as well as start/schedule sessions, etc. The Kanban board feature can be used to orchestrate agent driven workflows, and the agent itself can modify the Kanban lanes. Basically the agent can do all the same things you use the UI for.
Example prompt: “schedule a session using Opus 5 with max thinking for every feature in the PRD on the canvas. For every session make the prompt instruct the agent to review end to end test coverage for the feature and create a report on the canvas for test gaps that you find. Schedule the sessions overnight tonight spread evenly throughout the night, and have each session set to retry upon token exhaustion.”
I think it’s more meta than other meta-harnesses, but I’m biased because it’s my pet project.
I've tested Omnigent superficially, attracted to its thinking around policy, governance, sandboxing, and ui. But it's still alpha at present. I forked its Polly model and got working a somewhat more complex multiagent workflow that I've also modeled in Sandcastle and Gas City but the agent broke after the next update which I would have needed to patch to maintain functionality. Subjectively I also noticed individual models seemed to be performing somewhat worse when wrapped in the platform's framework, presumably due to the extra context introduced (token use was measurably higher). Promising project that I'll revisit when it's further along and I do not doubt the outcomes Databricks claims in committedly dogfooding it.
Double-harnessing distortions seems like a weak point. I wonder if it’s just a temporary measure and long term it’s about writing custom harness going straight to inference APIs across all models.
I've been using it for a week or so. The main draw for me is that I can keep my sessions in one database regardless of the model/provider I use. The webapp can access everything remotely, which is convenient when I'm on my phone.
I haven't gotten a chance to test the multi-agent capabilities, but the DeepSeek Flash prices are so low that I probably will soon.
These seem like the obvious tweaks akin to "using a cheaper hosting platform". I think the real savings come from careful context control for programmatic agents, careful tool awareness and usage to reduce thrashing, distilling workflows into deterministic processes and, moat importantly, adding friction and boundaries for non-technical users who tend to burn tokens making insane asks like "analyze all documents and give me a summary".
Reads like an add to Omnigent or whatever harness (wait it’s meta harness?.
How can Smart Router achieve higher task completion rate compared to any of the base models if all it does is dynamically switch base models based on cost??
not sure about the use of exponential and efficiency frontier here, these have formal sides to them but seem to be used rather inflationary and colloquially.
Appreciate the detail in this and the previous post on creating internal benchmarks!
Have you all attempted finetuning smaller OSS models on your repos for coding?
We do this for a lot of our customers (fine tuned to save cost when inference volume is high). Right now for internal coding we are using off-the-shelf models but we are considering fine tuning as well to squeeze more efficiency out.
I think there is a lot of dev cope in this thread.
My workflow is very simple:
1. develop requirements for code change
2. take manual notes for implementation, maybe use LLM for some discovery/investigation
3. present notes to frontier LLM
4. develop implementation plan (bulk of work)
5. let LLM rip
6. review diff, manually fixing/refactoring code as necessary, sometimes prompting for revisions
7. get automated LLM review
8. get human review
this reliably produces the work of 2-3 pre-AI senior engineers with a lower bug rate, equivalent performance, robust edge-case consideration, etc.
Does the LLM produce over-engineered solutions? All the time. I stop it from doing that, or manually fix it myself.
Does the LLM always adhere to the best system design? No, not at all. I often have to guide its design into a better, north-star aligned one.
I don't just sit in front of my terminal and say, "Ok Claude, build the app." It is a very iterative process, and not without its potential pitfalls.
But it is very, very productive.
This has been my experience as well. The best model I have access to right now is Opus 4.8. It's really good at fixing bugs in an established architecture or adding a similar feature, but it's absolutely mid tier at putting together a simple effective architecture for even common software problems. I can tell that functional programming practices are not a large part of it's training material.
Fortunately I’m at a startup with basically infinite Fable 5, and it is legitimately a huge step forward over Opus 4.8. Would recommend. Hopefully some open weight models catch up to its capabilities soon.
“use lower cost models”
“use price controls”
Truly revolutionary stuff.
how do any of these routing approaches handle kv cache misses? Devin Fusion is the only one that explicitly addresses this, though it does so by switching models during compaction (not sure this isn't still a cache miss though)
We're going to do a followup blog detailing our routing approach soon! In short, the router takes in the task description and infers what models and harnesses are available and makes a recommendation up-front. So essentially the routing decision is made when the harness + model is kicked off and it's only changed halfway through if there's a major delta in complexity from the initial judgment. Therefore, most of the time the cache is maintained just as it would be before (this is the advantage of having a meta-harness that is actually planning all the sub-agents centrally)
Maintaining the cache is extremely, extremely important, so we're iterating fast but that's a major factor we track in the router's development. Couple things I'd look at:
1. The cache is generally reset after a compaction - this is the best time to make a switch if you want.
2. In many cases, the max duration of a cache is 1h, so if a session is being resumed after a long time, that's also a good time to re-assess the complexity.
We're iterating fast here and learning a lot! Definitely a lot to think about it in this area.
The kv cache is wiped as soon as you get your answer, cloud hosts are not going to hold the GPU memory for your entire session. You're probably referring to some agent level cache
good engineer + llm = good engineer.
bad engineer + llm = bad engineer.
is this opensource or have to buy from Databricks?
It seems Databricks open-sourced it a while ago:
https://www.databricks.com/blog/introducing-omnigent-meta-ha...
https://github.com/omnigent-ai/omnigent
Omniagent looks quite similar to OpenRouter (https://openrouter.ai/)
Omnigent and OpenRouter are different in the sense that OpenRouter is where you can go to call the actual model but Omnigent is intended to be the place where you go describe the high level task to be done, and work is farmed out to various harnesses and models. Those sandboxes can themselves be using OpenRouter for capacity!
We're calling the layer coordinating harnesses "meta-harness'
Omnigent seems to compete more against Orca https://github.com/stablyai/orca They both went to be the Agent IDE layer, where you come with your tasks and everything is taken care of. I've been using Orca for a handful of tasks and have been largely enjoying it. My default barebones workflow is ghostty + zmx on ssh connections.
These tools casually like to claim they are orchestrators, but unfortunately, none of them are.
Ultimately, Databricks wants your enterprise on their platform. I dont think they particularly care about open source or the little guy.
So did we. I just asked my team to get personal accounts that I reimburse them for. It’s just a golden age loop though, the gravy train can’t go on forever unless we start building out thousands of data centers and associated renewable energy.
First the mofos force you to use AI then they become stingy about it.
An AI-edited post by the way.
Well yes, first hit is free.
Probably coulda got every dev a local model for how much they spent; what a brialliant set of economists
Really? Because removing it from my company has saved us over 2 million a year and we were able to speed up processing. The chargeback model for databricks is predatory at best.
What did you move to and what type of workload, if I may ask?
I think you’ve misunderstood the article. It’s about how Databricks reduced their own costs, not about how adopting Databricks will reduce anyone else’s costs.
Yawn. Databricks and their half baked overly expensive platform.
Too bad their AI query generation is next to useless.
Quit cold turkey and you can drive down AI coding spend 100%.